A survey is a request for someone's time, and the size of that request decides whether you get an answer. A long form emailed a week after the event is a large request, answered by a small and unrepresentative group. One question asked inside the product, moments after the experience it refers to, is a small request answered by people who still remember what happened. Everything else about survey design follows from that difference.
Choose the answer format before you write the question
Net Promoter Score asks how likely someone is to recommend you on a scale from zero to ten. It measures a relationship rather than an event, which makes it useful as a trend across quarters and close to meaningless as a reaction to a single interaction. It needs a steady flow of answers before movement means anything, so it belongs with your established users and a schedule you keep to, not with a launch week you want to feel good about.
A satisfaction question is the opposite: it is about one specific experience, asked immediately after it. Was the setup clear? Was the conversation helpful? One click, one moment, one thing to improve. Because the question is narrow, the answer is easy to act on — you know precisely which step the person was rating, and the people who rate it poorly are the ones to talk to next.
Multiple choice is for prioritisation rather than measurement. Between two and eight options, asked of people who have enough experience to have a preference, is how you find out which improvement matters most to the customers you already have. It is also the format most often misused: if the answers will not change what you build next quarter, the question is costing attention you will want later.
Timing decides the response rate
Timing is the biggest single factor in response rate, and it comes down to two settings. The page rule decides where the question appears, and the delay decides how long someone has been there. A satisfaction question on the settings pages a few seconds after a configuration is finished catches the experience while it is still fresh; the same question on the home page a fortnight later is asking people to reconstruct a memory.
Recency also protects you from the wrong answer. Ask during a task and you interrupt it, and the score partly reflects the interruption. Ask far too late and you measure how the last two weeks went rather than the thing you were asking about. The moment worth aiming for is just after a natural end: a finished setup, a resolved conversation, a completed order.
Ask the right people, and not too often
Who you ask matters as much as when. A loyalty question put to an anonymous visitor who arrived four minutes ago measures nothing. Restricting a relationship survey to existing users, and a satisfaction question to the people who have just been through the relevant step, gives you a smaller number of answers that actually support a decision.
Frequency is the other half of that discipline. Each person receives a given survey once, which prevents the same question following someone around, but nothing prevents three separate surveys running at once and consuming the same goodwill. Treat responses as a budget: one question in the field at a time, with a deliberate gap before the next one, will outperform a permanent stream of small requests.
Read answers next to the conversation
Answers are worth more when they sit beside the person who gave them. A survey response recorded against a contact can be read next to their conversation history, their plan and the pages they were on, which turns a low score from a statistic into a customer you can follow up with by name — the single most reliable way to learn what the number actually meant.
Close the loop and say what changed. Reply to the people who left a poor score, group the free-form answers into themes rather than counting them, and take the top theme to the roadmap. When the change ships, announce it to the same audience you asked. A team that visibly acts on answers keeps its response rate; a team that only tracks the score watches it decline and learns nothing about why.
Run one survey properly instead of three at once
Write the decision before the question. Put in one sentence what you will do differently depending on the answer — change the setup flow, reprioritise the roadmap, call the customers who scored you poorly. A survey with no decision attached produces a number that gets reported quarterly and acted on never, and it spends the attention you would rather have available when a real question comes up.
Then choose the format the decision requires. A trend you want to watch over quarters is a nought-to-ten score aimed at established users. A step you suspect is confusing is a satisfaction question on that step's page, seconds after it ends. A choice between three things you could build next is a multiple choice question put to the people who would use them. Getting this pairing right matters more than the wording.
Publish one survey, leave it alone for a fortnight, and read it in full. Look at the distribution rather than the average, read every comment, and check that the answers are coming from the audience you intended. Then act on one theme, announce the change, and only afterwards put the next question in the field. Slower rotation gives you higher response rates and cleaner comparisons than running everything at once.
Frequently asked question
How many answers do I need before an NPS score means anything?
There is no universal threshold, but a score built on a handful of responses moves so much with each new answer that the movement is noise. Small teams get more from the distribution and the comments than from the headline number: a shift from mostly high scores to a cluster of low ones is a real signal at any sample size, while a two-point change on a few dozen answers usually is not. Keep the audience, the moment and the wording constant between rounds, because changing any of them makes the comparison meaningless regardless of volume.